Meiming Shen

Design of an I/O balancing file system on Web server clusters

The paper presents the design of an I/O balancing file system on Web server clusters. High-speed networks allow machines to access remote data nearly as quickly as they can access local data. This trend motivates the introduction of an I/O balancing mechanism into cluster file systems. Two effective approaches are issued in the paper: file migration and unique cache. The former is to migrate the files from congested nodes to those uncongested ones to balance the I/O load among the cluster system. And the latter is to integrate the memory and disk cache to form a global memory-disk file cache. It extracts the ideas of cooperative cache and file replication to improve the file system performance. Also unique cache balances the I/O load at the file cache level. We can conclude that TH-CluFS is a suitable file system for Web server clusters.

An effective file migration algorithm in cluster file systems

I/O balancing is an attractive problem in distributed systems. With the fast growth of high-speed network technology, file migration is becoming an effective way to solve the problem. We present an effective file migration algorithm in our cluster file system - TH-CluFS. Our algorithm moves files from congested nodes to relatively uncongested nodes according to the information of current system states. The algorithm makes each file migration beneficial to the mean response time of file requests. Moreover we use simulation to demonstrate that file migration can improve system overall performance greatly in cluster file systems. Our results indicate that the cluster system with file migration algorithm not only achieves higher scalability and more powerful processing capability, but also makes the effect of network traffic reduced about 12%.

Grid-Based Biological Computation Service Environment

This paper puts forward a novel concept of “Grid-based biological computation environment”. In this environment, distributed, heterogeneous computer nodes, as long as they installed biological computation software and are willing to share them for public use, can be aggregated seamlessly into a powerful computational Grid to provide on-demand computation services. Just with the widely available Web browser, biology researchers can conveniently perform highly customized computation jobs. To date, a prototype system is developed and a number of pragmatic biological applications are exploited.

Extending the Power of Server Based Computing

In this paper, we review the evolution of computing, especially the development of thin client computing. Then, our study is focused on one of the most popular thin-client computing mode: server based computing. After an analysis of the problems existing in server based computing, a revised server based computing model is given, which combines the advantages of other thin-client model. As proof of our concept, our ideas are implemented in detail on our embedded platform: THUMP-NC. Some open source software, including rdesktop and VNC, are used in our implementations.

Event chain clocks for performance debugging in parallel and distributed systems

In this paper, the Event Chain Clock synchronization algorithm is presented. This algorithm can maintain a global physical clock that reflects both the partial order and the elapsed time of all events occurred. This algorithm, which repeats some basic operations, has good astringency, and is suitable for parallel program performance debugging.

PREFETCHING AND CACHING METADATA IN A DISTRIBUTED NFS SERVER

The paper describes a name and attribute cache(NACache) in a distributed NFS server. The distributed NFS server provides the identical NFS service as the common NFS server and no modification to NFS cli ent software. The NACache stores the metadata information of the distributed server and employs the prefetching approach to optimize the performance of metadata operations. It retrieves all file attributes and names in a single request and stores them in the NACache. In the paper, the design and implementation of the NACache in the distributed NFS server is presented. And the performance of our system is evaluated via the Andrew Benchmark. According to the experimental results, we conclude that the NACache can improve the system performance substantially and the perfetching approach can increase the hit rate in the cold cache state significantly.

Data I/O Optimization in Storage Systems

One of the most effective ways to improve the I/O performance of a storage system is to enhance the hard disk’s read/write ability. We used an I/O processing node in the storage network to optimize data organization and I/O performance. By analyzing existing algorithms and different requirements for read and write operations, we designed an improved optimizing algorithm to schedule disk I/O requests. It selects the closest request in queue to process first, and uses an EW mechanism to modify write locations. Typically, the algorithm can reduce a disk’s average response time by about 15%-17%. This paper also presents an EW stripe and copy algorithm that can improve I/O performance using parallel disk accesses, and enhance reliability by data duplication. With one copy preserved, it can reduce the response time by about 30%.

Improving the Objects Set Availability in the P2P Environment by Multiple Groups

The availability is often used to measure how well a system can provide the service continuously. The measures of availability for a single computer and for multiple computers are very different. In this paper, we discuss the objects set availability in the distributed environment especially in the case of P2P environment. The objects set availability is defined and formalized. In the P2P environment, the hosts are divided into groups by their availability values. Objects are evenly distributed among groups. The experiments show that such simple but practical method has good efficiency in achieving high objects set availability.

A class-based garbage collector for network server applications

As garbage collection is a key feature in modern application environments, performance issues are mostly important for the efficiency of applications. Especially in the case of network service applications, the server should run for as long a time as possible, during which many objects will be frequently created and destroyed. The garbage collection then becomes one of the critical factors of the servers response time. In this article, we present a class-based garbage collector. It classifies objects according to their class, i.e. the type information, and groups those objects that belongs to the same type into the same group. By doing so, we can minimize the object creation/collection time and can also improve the gradualness of the applications that run on the network server.